SLAM - PAPER - 1

Matching-range-constrained real-time loop closure detection with CNNs features

具有匹配范围约束的带重复特征的实时环路闭包检测

  • abstract

    • loop closure detection 闭环检测LCD
    • DCNNs
    • Some researchers

      • pre-trained CNNs model
      • generating an image representation
      • appropriate for visual LCD in SLAM
    • Differences and Challenges between Simple Computer Vision & Robotic Application

      • adjacent images more resemblance
      • real-time performance
    • in this paper

      • making use of the feature generated by CNNs layers

        • to implement LCD in real environment
      • the first problem

        • provide a value to limit the matching range of images
      • better results

      • improve the real-time performance

        • using an efficient feature compression method
  • background

    • SLAM algorithm aims

      • map an unknown environment
      • while simultaneously localizing the robot
    • LCD

      • determine whether a robot is back to previously visited location
      • correcting the accumulate error is critical for building a consistent map
      • One of the most essential techniques in SLAM
    • develop a LCD algorithm

      • one class of popular and successful techniques

        • based on Matching the current view of the robot with
        • Those Corresponding to previously visited locations in the robot map
        • OTHER

          • image matching problem
        • STEPS

          • image description
          • similarity measurement
        • The state-of-the-art Algorithms

          • image description

            • BoW : bag-of-words model

              • clusters the visual feature descriptors in images

                • visual features (Success)

                  • SIFT
                  • Surf
              • builds the dictionary

              • find the corresponding words of image
          • similarity measurement

      • Challenges still remain

        • in dynamical and large-scale environment
        • long period of time

          • day
          • week
          • months
        • dramatic condition change

        • viewpoint change over time
        • bad news

          • the hand-craft methods can not deal with these situations very well
        • good news

          • in recent ML and CV conference
          • the features generated by convolutional neural networks outperform well in visual recognition classification and detection applications
      • advantage of CNNs

        • it has been demonstrated to be versatile and transferable
        • even though they were trained on a very specific target task
        • they can be successfully used to solving different problems
        • and may outperform traditional hand-craft features
    • Two Challenges appear while we use these features generated by CNNs in pratical environment

      • Firstly , the adjacent images in the data-set of LCD might have more resemblance than the images that really form the loop closure

        • #?为什么相邻图像比真实构成闭环检测的图像更相似?
        • the algorithm tends to identify the adjacent images as loop closure
      • Secondly , the feature matching is computationally intensive , because the dimension of features generated by CNNs may be very large .

LCD have to compare the current image to large amount of pre-captured images

        - 太大的计算量不利于实时性

- in this paper

    - two solution

        - firstly

            - provide matching range of candidate images

                - #将匹配到一张图片,变成匹配到相似的图片范围里去

        - secondly

            - a efficient feature compression method

                - #有效在于,通过压缩了CNN层得到的图像特征,这样处理的图像变小,处理速度快一点就能提高实时性, (临界性能减小)

            - to reduce dimension of feature generated by CNNs

ViNS-Mono

遥感影像道路提取研究

基本思路

  • 首先利用各种特征提取方法提取有用特征
  • 应用各种方法找出满足道路特征的道路
  • 最后对道路提取结果进行后期优化处理得到最终道路提取结果

目前的方法

  • 特征提取
  • 道路提取

    • 1、同时包含道路拓扑结构信息和宽度信息的提取
    • 2、只提取出道路中心线的拓扑结构信息

Long Range Traversable Region Detection Based on Superpixels

Clustering for Mobile Robots
可达性

abstract

  • — Traversable region detection
  • i传统方法缺点

    • only short range traversable regions can be detected
    • 原因

      • the limited image resolution and baseline of stereo vision
  • 本文方法

    • detect long range traversable regions without using any supervised or self-supervised learning process
    • Superpixels clustering algorithm
    • superpixels are clustered using an improved spectral clustering algorithm to segment the image
    • integrating short range traversable region detection : u-v-disparity

      • then the traversable region can be extended to long range naturally
  • result

    • works well in different outdoor environments
    • detecting range can be improved greatly

Introduction

  • 可达性

    • regions that do not contain geometric obstacles
  • 采集信息

    • ultrasonicsensor
    • stereo vision

      • measure the ranges to objects
      • by calculating disparities between stereo images
    • laser scanners

  • key

    • After acquiring the disparities ,traversable regions or obstacles can be detected robustly and efficiently
      (using a series of approaches based on u-v-disparties )
  • V-disparity

    • Aim

      • detect Obstacles
    • (u,v)

      • 坐标
    • ways

      • by accumulating pixels with the same disparity value d in each row , a v-disparity image (d,v) was build
      • perpendicular obstacles can be mapped to vertical lines

        • pixel intensity represents the width of obstacles
      • the traversable region modelled as a succession of planes can be projected as slanted line segment